rootonchair/MiniMax-H3-nunchaku-lite-nvfp4

MiniMax-H3 transformer, Nunchaku Lite nvfp4 (data-free SVDQuant)

The 30.8B MiniMax-H3 video+audio transformer, quantized with data-free SVDQuant (weight-span smoothing + rank-32 SVD low-rank branch + NVFP4 group quantization, group size 16, 362 quantized linears) - no calibration data. Packed in the Nunchaku Lite kernel layout; loads through diffusers' pre-quantized fast path.

This is the transformer component only - pair it with the other components from MiniMaxAI/MiniMax-H3:

import torch
from diffusers import MiniMaxH3Transformer3DModel

transformer = MiniMaxH3Transformer3DModel.from_pretrained(
    "rootonchair/MiniMax-H3-nunchaku-lite-nvfp4", dtype=torch.bfloat16, device_map="cuda",
)

Requires the kernels package, DIFFUSERS_TRUST_REMOTE_KERNELS=true, and a Blackwell-or-newer NVIDIA GPU (NVFP4 kernels need sm_120; PyTorch >= 2.7 with CUDA >= 12.8).